NEKO//AGENT Research
🐾
Evidence-led research for safe, tool-using agents
NEKO//AGENT — AGENT SYSTEMS
Reasoning with actions, learned tool use, reflection, agent-computer interfaces, and executable agent evaluation.
Note Public research Space with deterministic trace evaluation and explicit sandbox boundaries; not a live autonomous agent.
ReAct: Synergizing Reasoning and Acting in Language Models
Paper • 2210.03629 • Published • 38Note Interleaves reasoning traces with actions against external environments.
Toolformer: Language Models Can Teach Themselves to Use Tools
Paper • 2302.04761 • Published • 12Note Studies language models learning when and how to invoke external tools.
Reflexion: Language Agents with Verbal Reinforcement Learning
Paper • 2303.11366 • Published • 9Note Language-agent reflection and feedback through episodic textual memory.
SWE-bench: Can Language Models Resolve Real-World GitHub Issues?
Paper • 2310.06770 • Published • 13Note Executable evaluation of systems resolving real software issues.
SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering
Paper • 2405.15793 • Published • 7Note Studies how agent-computer interface design changes software-agent performance.
princeton-nlp/SWE-bench
Viewer • Updated • 21.5k • 91.9k • 149Note Executable software-engineering task dataset; use exact harness and repository revisions.
gaia-benchmark/GAIA
Viewer • Updated • 932 • 11.5k • 882Note Gated assistant benchmark. Preserve its current non-resharing and leakage controls.